For example, to reduce noise components included in an image signal, various methods have been proposed so far. Particularly, one of the simplest methods having a large noise reduction effect is a method using a low-pass filter (hereafter referred to as LPF). The LPF is a device for transmitting only signals having components lower than a reference frequency. That is, by inputting a signal whose frequencies change to the LPF and observing the amplitude of an output signal, a characteristic is obtained that a component at a higher frequency lowers in level.
However, when viewed from a different point, the LPF uses the average value of a watched pixel and adjacent pixels around the watched pixel as a new value of the watched pixel. That is, in the case of this method, signal levels of watched pixels strongly correlated with peripheral pixels are not greatly changed in their values even if the levels are averaged. However, random noise components having no correlation are averaged with noise components included in peripheral pixels and thereby, the value of the component is approached to “0”.
Therefore, when using the above LPF, the noise suppression effect increases as the search area of peripheral pixels is widened. However, in the case of the averaging operation with peripheral pixels by the LPF, image edge information is reduced similarly to noises and resultantly, the whole image becomes blurry though noises are decreased and a disadvantage occurs that the image quality is deteriorated. Therefore, an LPF serving as noise reduction means is not generally used.
To solve the disadvantage of the LPF, the so-called ε-filter is disclosed (refer to Journal of Institute of Electronics, Information, and Communication Engineers Vol. 77 No. 8, pp. 844–852, April, 1994, Kaoru Arakawa “Nonlinear Digital Filter and Its Application”). That is, in the case of the ε-filter disclosed in this document, when averaging a watched pixel and peripheral pixels, it is first determined whether the peripheral pixels has a correlation with the watched pixel.
Specifically, by setting a certain reference level θ, levels of the peripheral pixels are incorporated into averaging factors when the levels are included in the range of ±θ of the level of the watched pixel but they are not incorporated into averaging factors if they are not included in the range of ±θ. Thus, whether to incorporate all peripheral factors into averaging factors is searched and a new value of the watched pixel is obtained by the averaging operations with the watched pixel and the peripheral pixels which are regarded as operation objects.
Therefore, even if an image edge enters a search area, when the levels of pixels constituting the edge exceeds the range of ±θ of the level of the watched pixel, the edge is not regarded as an operation object, for example, it never happens that an image becomes blurry due to pixels constituting the edge being included in averaging. That is, with the ε-filter, it is possible to suppress only noise components while leaving an image edge as it is by properly selecting the value of he reference level θ.
Moreover, an actual circuit configuration of the ε-filter is described below by using FIG. 5. In FIG. 5, the diagram 1 shows a certain one point in an image area and imaged states of a watched pixel o and its peripheral pixels a, b, c, d, e, f, g, and h. Moreover, when substituting level values of these pixels with the same notation as a to h and o, the level values a to h of these peripheral pixels are supplied to a selection circuit 2. Moreover, the value of the above reference level θ and the level value o of the watched pixel are input to the selection circuit 2.
In the selection circuit 2, the absolute value (|a−o|) of the difference between the level value a of the peripheral pixel a and the level value o of the watched pixel o is first calculated and the absolute value of the difference is compared with the reference level θ. Then, when the absolute value of the above difference is smaller than the value of the reference level θ, the level value a is output to an output port 3. Moreover, when the absolute value of the difference is larger than the value of the reference level θ, the level value a is not output to the output port 3 but the value “0” is output. Furthermore, the same calculations are applied to level values b to h of other peripheral pixels b to h.
Therefore, eight output ports 3 equal to the number of peripheral pixels, for example, are provided for the selection circuit 2, and the level values a to h are output to the output ports 3 when the absolute value of the above difference is smaller than the value of the reference level θ and the value “0” is output to the ports 3 when the absolute value of the difference is larger than the value of the reference level θ. Moreover, an output port 4 is provided for the selection circuit 2 and a value obtained by adding “1” to the number of the output ports 3 to which the above level values a to h are output is output to the output port 4.
That is, level values a to h are output from the output ports 3 of the selection circuit 2 when absolute values of differences between a watched pixel and peripheral pixels are all smaller than the value of the reference level θ and the value “9” is output to the output port 4. Moreover, when absolute values of differences between the watched pixel and peripheral pixels are all larger than the value of the reference level θ, the value “0” is output from all output ports 3 and the value “1” is output from the output port 4.
Outputs of the output ports 3 of the selection circuit 2 and the level value o of the watched pixel o are supplied to an adder 5 and a value selected by the output port 6 of the adder 5 is supplied to a divider 7. Moreover, a value outputted from the output port 4 of the selection circuit 2 is supplied to the divider 7. Then, in the divider 7, a value outputted from the output port 6 of the adder 5 is divided by a value outputted from the output port 4 of the selection circuit 2 and the value of the above operation result is output by an output port 8.
A certain reference level θ is set, and levels of the peripheral pixels are incorporated into averaging factors when the levels are included in the range of ±θ of the level of a watched pixel but the levels are not incorporated into averaging factors when they are not included in the range and then, whether to incorporate all peripheral pixels into averaging factors is searched and only peripheral pixels to be incorporated as averaging factors are regarded as operation objects and as a result, a new value of a watched pixel obtained through the averaging operation with the watched pixel is output to the output port 8.
A specific circuit configuration of the selection circuit 2 of the above device is similar to the configuration shown in FIG. 6. That is, in FIG. 6, for example, eight comparators 20 equal to the number of the above peripheral pixels are obtained. Level values a to h of the above peripheral pixels, the level value o of the watched pixel, and the value of the reference level θ are input to the comparators 20. Then, each comparator 20 outputs the value “1” when the absolute value of the difference between a peripheral pixel and the watched pixel is smaller than the value of the reference level θ.
Moreover, a signal output from each of the comparators 20 is supplied to an AND gate 21. Furthermore, level values a to h of peripheral pixels are supplied to the AND gate 21 and corresponding one of the level values a to h of peripheral pixels is output to the output ports 3 through the AND gate 21 when a signal output from each of the above comparators 20 is equal to “1”. Furthermore, signals output from the comparators 20 are supplied to an adder 22. Furthermore, an addition output of the adder 22 is supplied to an adder 23 and the value “1” is added and output to the output port 4.
Thereby, in the case of this circuit configuration, level values a to h of peripheral pixels are output through the AND gate 21 when absolute values of differences between level values a to h and the level value o of the watched pixel are smaller than the value of the reference level θ. Moreover, the value “0” is output when absolute values of the differences are larger than the value of the reference level θ. Furthermore, a value obtained by adding “1” to the number of level values a to h output to the output ports 3 through the above AND gate 21 is output to the output port 4.
Thus, the selection circuit 2 outputs level values a to h when absolute values of the above differences are smaller than the value of the reference level θ and a value obtained by adding “1” to the number of output level values a to h. Moreover, the level values a to h and the level value o of the watched pixel are added and the addition value is divided by a value obtained by adding “1” to the number of output level values a to h. Thereby, the averaging operation is applied to only pixels regarded as averaging factors and a new value of the watched pixel is derived.
Thus, in the case of the above ε-filter, it is possible to effectively reduce noises while preserving an image edges. However, even when using the ε-filter, because basic processing is performed by an LPF, image details having a high-frequency component and a small amplitude disappear. Meanwhile, when an object is flat, noises easily become conspicuous and the ε-filter has a large effect. However, when there are many high-frequency components, the effect of the ε-filter is small and noises do not easily become conspicuous.
Therefore, when an object has many high-frequency components, the ε-filter is turned off. However, it is alternative whether to perform signal processing by the ε-filter or not. In this case, a case occurs that a state in which processing is performed or not performed is not necessarily the optimum image processing. That is, noises increase when no operation is performed or image details disappear when an operation is performed depending on the content of an object.